temporal depth video enhancement based on intrinsic static ... · intrinsic static structure icip,...
TRANSCRIPT
Temporal Depth Video Enhancement Based On
Intrinsic Static Structure
Lu Sheng, King Ngi Ngan & Songnan Li
The Chinese University of Hong Kong
Outline2
Introduction & Related Work
Intrinsic Static Structure
Temporal Depth Video Enhancement
Experimental Results
Conclusion & Future Work
ICIP, Oct. 27-30, 2014, Paris, France
Introduction
In a raw depth video capturing a natural scene
Various complex and even unpredictable dynamic contents
As well as spatial and temporal artifacts
Artifacts in the depth video
Both in Spatial and temporal domain
Errors are caused by
Low resolution
Noise & outliers
Occlusions, material absorption & reflection
Distortion around object boundaries, and etc.
ICIP, Oct. 27-30, 2014, Paris, France
3
Raw Kinect video
Color-coded Raw TOF video
Introduction
After the spatial enhancement
Reduce artifacts in spatial domain
But not enough to deal with temporal flickering
No temporal consistency
Aggravate flickering artifacts
ICIP, Oct. 27-30, 2014, Paris, France
4
Raw Kinect video
Spatial JBF [1]
[1] J. Kopf et al, Joint bilateral upsampling, in ACM ToG., 2007.
Introduction
After the spatial enhancement
Reduce artifacts in spatial domain
But not enough to deal with temporal flickering
No temporal consistency
Aggravate flickering artifacts
After a conventional spatiotemporal enhancement
Still contain temporal flickering
Distort the necessary depth variation along dynamic objects
How to eliminate the temporal flickering while do not distort the necessary depth variation along the dynamic objects?
ICIP, Oct. 27-30, 2014, Paris, France
5
Coherent spatiotemporal JBF [2]
Spatial JBF
[2] N. Richardt, et al, Coherent spatiotemporal filtering, upsampling and rendering of RGBZ videos, Computer Graphics Forum, 2012
Related Work
Spatial Enhancement
Global optimization
J. Diebel et al., NIPS 2006
J. Yang et al., ECCV 2012
J. Park et al., ICCV 2011
and etc.
High dimensional Filtering
J. Kopf et al., TOG 2007
J. Dolson et al., CVPR 2010
B. Huhle et al., CVIU 2010
and etc.
Patch Match and etc.
J. Lu et al., CVPR 2013
L. Sheng et al., ICIP 2013, and etc.
6
ICIP, Oct. 27-30, 2014, Paris, France
Temporal Enhancement
Median filter
S. Matyunin et al., WSCG 2011
norm
S. H. Chan et al., TIP 2011
Texture consistency
Dongbo Min et al., TIP 2012
Deliang Fu et al., PCS 2010
Both texture and depth
N. A. Dodgson et al., CGF 2012
Model the scene
J. Shen et al., TIP 2013
J. Shen et al., CVPR 2013
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
7
A moving object
A static object
The static background
Kinect or another depth camera
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
8
A moving object
A static object
The static background
Kinect or another depth camera
Captured depth map
Intrinsic Static Structure
Definition
A manifold lies on or behind the input depth map
Contains static structure of the captured scene
ICIP, Oct. 27-30, 2014, Paris, France
9
A moving object
A static object
The static background
Kinect or another depth camera
Intrinsic static structure
Intrinsic Static Structure
Observation
Moving objects stay in its front
Static regions or visible background area are fused into it
ICIP, Oct. 27-30, 2014, Paris, France
10
A moving object
A static object
The static background
Kinect or another depth camera
Intrinsic static structure
Intrinsic Static Structure
How to estimate the intrinsic static structure under an on-line fashion?
Temporal enhancement of the depth video?
11
ICIP, Oct. 27-30, 2014, Paris, France
A probabilistic generative model
An online variational probabilistic update scheme with
A robust exclusion of the dynamic foreground objects
the inclusion of the once occluded background
Robust static region detection by the intrinsic static structure
Enhance the static region of input depth frame
Dynamic objects own limited temporal consistency, we just omit it
Fusing the static region with the estimated intrinsic static structure
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
12
Camera center
Line of sight
Current intrinsic static structure
Behind the structure
Before the structure
is an indicate function that is equal to 1, when input argument is true and 0 vice visa
A Probabilistic Generative Model
If incoming depth belongs to
State-I: the intrinsic static structure
State-II: outliers in the front or moving objects
State-III: outliers rearward or revealedbackground
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
13
Camera center
Line of sight
Current intrinsic static structure
State-I
is an indicate function that is equal to 1, when input argument is true and 0 vice visa
A Probabilistic Generative Model
If incoming depth belongs to
State-I: the intrinsic static structure
State-II: outliers in the front or moving objects
State-III: outliers rearward or revealedbackground
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
14
Camera center
Line of sight
Current intrinsic static structure
State-II
is an indicate function that is equal to 1, when input argument is true and 0 vice visa
A Probabilistic Generative Model
If incoming depth belongs to
State-I: the intrinsic static structure
State-II: outliers in the front or moving objects
State-III: outliers rearward or revealedbackground
Intrinsic Static Structure
ICIP, Oct. 27-30, 2014, Paris, France
15
Camera center
Line of sight
Current intrinsic static structure
State-III
is an indicate function that is equal to 1, when input argument is true and 0 vice visa
A Probabilistic Generative Model
If incoming depth belongs to
State-I: the intrinsic static structure
State-II: outliers in the front or moving objects
State-III: outliers rearward or revealedbackground
Intrinsic Static Structure
A Probabilistic Generative Model
The likelihood of w.r.t. the given intrinsic static structure
Gaussian prior over
Dirichlet prior over the frequency of each state
ICIP, Oct. 27-30, 2014, Paris, France
16
Camera center
Current intrinsic static structure
Intrinsic Static Structure
A Probabilistic Generative Model
The posterior
ICIP, Oct. 27-30, 2014, Paris, France
17
Camera center
is the set of previous depth samples
is the set of current samples
On-line Update Scheme
A Probabilistic Generative Model
The posterior
ICIP, Oct. 27-30, 2014, Paris, France
18
Camera center
Updated intrinsic static structure
is the set of previous depth samples
is the set of current samples
If the input frame only contains the static scene and outliers, the updated intrinsic static structure will be governed by the posterior, and we have
Its probable depth is
The reliability of the model is
Variational approximation for efficiency
On-line Update Scheme
ICIP, Oct. 27-30, 2014, Paris, France
19
Variational approximation for a 1D depth signal
a) the probable depth values of the intrinsic static structure versus the raw depth signals
b) The confidence interval of the intrinsic static structure
c) The portions of three states
d) The approximated date distribution versus the histogram of the raw data
The nature of the input data samples is still well captured by the variational approximation
On-line Update Scheme
ICIP, Oct. 27-30, 2014, Paris, France
20
If the input frame contains the dynamic content, the posterior w.r.t. each state indicates the criterion to update the intrinsic static structure
State-I: fuse the new samples into the current modelState-II: leave the current model unchangedState-III: have a risk that current model is incorrect,
might require re-initialization
Camera center
State-III
State-I
State-II
Updated intrinsic static structure
Results after variational approximation are valid for State-I and State-II
Only need an additional criterion for State-III
On-line Update Scheme
Re-initialization to avoid incorrect estimation when
Ratio
It means the percentage of State-III is too large to provide enough confidence that current estimation is true
Increment
It means there are a successive frames that the states of input depth samples are State-III
Or texture consistency is violated at the locations of State-III
If RGB-D video is applicable. One step further, we can exploit the conditional random field to robustly find the re-initialized region
ICIP, Oct. 27-30, 2014, Paris, France
21
Camera center
State-III
State-I
State-II
Updated intrinsic static structure
On-line Update Scheme
ICIP, Oct. 27-30, 2014, Paris, France
22
Parameter initialization
When input depth sample is valid
No texture information is applicable
When input depth sample is valid, just copy the previous model
Temporal Depth Video Enhancement
Blend input depth frame/spatially enhanced depth frame with the estimated intrinsic static structure
ICIP, Oct. 27-30, 2014, Paris, France
23
Camera center
Updated intrinsic static structure
Temporal enhanced depth frame
The intrinsic static structure
The posterior of state-I
The temporally enhanced depth value
For RGB-D video, Both depth and texture information are helpful to fill depth holes
Inference the chance that the hole region belongs to the intrinsic static structure
Experimental Results
Intrinsic static structure estimation – 1D case
ICIP, Oct. 27-30, 2014, Paris, France
24
Experimental Results
Intrinsic static structure estimation – 2D case
ICIP, Oct. 27-30, 2014, Paris, France
25
Experimental Results
Temporal enhancement on synthesis static scenes
ICIP, Oct. 27-30, 2014, Paris, France
26
Input disparity map is contaminated by
Per-frame Running time comparison (MALTAB Platform)
Experimental Results
Temporal enhancement on static scenes
ICIP, Oct. 27-30, 2014, Paris, France
27
The reliability map indicates that most of the flat or smooth surfaces of the captured scene are of high reliability to be static
Around edges or highly slanted surfaces, the reliability map owns lower values
The reliability quantity
Experimental Results
Temporal enhancement on static scenes
ICIP, Oct. 27-30, 2014, Paris, France
28
Experimental Results
Temporal enhancement on dynamic scenes
ICIP, Oct. 27-30, 2014, Paris, France
29
Effectively and accurately find the moving objects according to the state classification, with the help of conditional random field and the registered color frame
Enhance the foreground region by Joint bilateral filter individually, fill in the holes in the static region by the estimated intrinsic static structure
Suppress noise and flickering artifacts at static regions
Experimental Results
Temporal enhancement on dynamic scenes
ICIP, Oct. 27-30, 2014, Paris, France
30
Effectively and accurately find the moving objects according to the state classification, with the help of conditional random field and the registered color frame
Less motion delay will occur
(c) Coherent spatiotemporal JBF (d) Spatiotemporal WMF (e) Ours
Experimental Results
ICIP, Oct. 27-30, 2014, Paris, France
31
(a) (b)
Spatiotemporal depth video enhancement
Conclusion & Future Work
Temporal flickering problem of conventional depth enhancement approaches
Introduction of the intrinsic static structures and its application to temporal enhancement of a depth video
Discuss the its potential to handle the static and dynamic scenes
Proper way to further combine spatial and temporal enhancement based on the intrinsic static structure into a whole framework
Research on a more general formulation of the static structure
How to deal with depth video captured by moving cameras, and etc.
ICIP, Oct. 27-30, 2014, Paris, France
32
33
Any Questions?
ICIP, Oct. 27-30, 2014, Paris, France
On-line Update Scheme
Variational Parameter Estimation
Factorize the posterior into independent Gaussian and Dirichlet distributions
The reliability of the model
The probable depth is
The posterior can be approximated by
ICIP, Oct. 27-30, 2014, Paris, France
34
Recursive estimation is possible!
On-line Update Scheme
Variational Parameter Estimation
Factorize the posterior into independent Gaussian and Dirichlet distributions
The posterior can be approximated by
Moment matching to estimate the hyperparameters
ICIP, Oct. 27-30, 2014, Paris, France
35
Closed-form
solutions!